How OpenAI’s Codex Is Merging Into ChatGPT to Power a “Super App” for 900 Million Users

Preface

After Anthropic released the Claude Code preview early last year, OpenAI realized that the smaller competitor had actually surpassed it in coding. Around the same time OpenAI formed a dedicated team to develop an AI coding tool.

Nearly 18 months later, OpenAI leaders—including Thibault Sottiaux, who heads the new team—concluded that Codex outperforms the flagship ChatGPT on many tasks.

Sottiaux told reporters that Codex excels at long‑running multi‑step tasks, using external tools, and writing code to manipulate complex spreadsheets.

OpenAI now positions Codex as a tool not only for programmers but also for a broad audience. A research report released on Tuesday describes Codex as becoming a productivity tool for everyone, helping people automate daily work, speed up workflows, and eliminate bottlenecks in modern knowledge work.

Plan to integrate ChatGPT, Codex and Atlas into a desktop “super app”

OpenAI announced that the previously disclosed plan to combine ChatGPT, Codex and the Atlas browser into a single desktop “super app” will launch in the coming weeks. Leadership said that by merging these applications they can focus on their most successful product—currently Codex.

In recent months OpenAI released models that handle large enterprise codebases better, helping Codex win more developers over, whereas previously most developers preferred Anthropic’s Claude Code.

Codex broke 5 million weekly active users at the end of May, after surpassing 4 million a month earlier and 3 million two weeks before that.

According to insiders, co‑founder and president Greg Brockman told staff last month that revenue from enterprise customers using Codex grew 50 % week‑over‑week, and CEO Sam Altman noted in an all‑hands meeting that overall usage is increasing by about 5 % per day.

Most of the 5 million weekly active users are paying customers, though a sizable fraction are still unpaid. Anthropic does not disclose user numbers for Claude Code and Claude Cowork, which are only available to paying users.

Anthropic’s revenue from these products may already exceed Codex’s current revenue. Claude Code’s popularity among developers helped Anthropic reach an annualized revenue of over $47 billion last month—a five‑fold increase from the start of the year. OpenAI disclosed $25 billion annualized revenue in March; recent reporting by The Information says OpenAI has now surpassed $30 billion.

By building the super app, OpenAI aims to bring Codex to more than 900 million consumer users and boost revenue, a move that is critical as both companies race toward IPOs to fund AI chip purchases and research hiring.

The integration of Codex with ChatGPT is also crucial for OpenAI’s goal of building agents that automate tasks for knowledge workers, similar to how Anthropic’s Claude Code enabled its broader Claude Cowork tool.

Coding is the key

Even before Anthropic released Claude Code in February 2025, OpenAI was worried about Claude’s coding abilities. In autumn 2024 internal metrics showed Claude leading OpenAI’s models, a concern because OpenAI has long believed AI‑assisted coding would accelerate research toward super‑intelligent AI.

In January 2025 OpenAI launched Operator, a browser‑based agent for tasks like travel planning or ordering supplies. However, leaders quickly realized that clicking in a browser was too slow and that writing code might be a more efficient way for AI to control a computer for tasks such as generating spreadsheets.

Engineer customers said OpenAI responded to Claude’s progress by developing models better at handling large enterprise codebases and, after launching the Codex desktop app in February, attracted more users.

A “startup within a startup”

A key factor was the creation of a dedicated Codex team that operates almost like a small startup inside OpenAI.

OpenAI product lead Alexander Embiricos noted that the team open‑sourced the code behind Codex, making it easier to gather feedback from users—a rarity in most companies.

Some OpenAI employees argue that Codex has an advantage because its model and product are developed by the same team, whereas the applications division is separate from the research organization.

Sottiaux said Codex’s superior “tool‑calling framework”—software that lets AI agents use tools on behalf of users—is a key reason for merging it with ChatGPT.

Reorganization: a unified organization

The success of the Codex team helped drive a January reorganization to bring product teams closer to the researchers who build the underlying models.

Last month OpenAI merged the teams responsible for ChatGPT, Codex and their APIs into a single organization led by Sottiaux, who now oversees “core products and platforms.”

Cloud vs. local: the toughest fusion

One of the biggest differences between Codex and ChatGPT is that ChatGPT runs in the cloud while Codex runs locally on the user’s device and can access local files. Sottiaux said his biggest current challenge is merging these two execution modes.

He must achieve this fusion in the upcoming super app. Embiricos said that in the short term OpenAI’s goal is to let users choose whether Codex or ChatGPT answers a query, making it clear which tasks Codex can now handle.

In the long term, the model will decide on its own whether to run on the device or in the cloud, removing the need for users to think about tool selection.

Developer reaction: Codex wins back developers

Codex’s resurgence is mainly because it has regained developer trust, and many developers now prefer Codex over Claude Code.

For example, Mitch Troyanovsky, co‑founder of New York AI‑accounting startup Basis, said their engineers hit the enterprise credit‑card limit and could not continue using Codex.

After engineers bombarded him on Slack, the usage limit was raised. They can still use Claude Code (which they pay for), but it lacks appeal; a developer even shared a GIF of a soldier handing a gun to a monkey to mock Claude Code.

Troyanovsky added that some engineers who frequently use Claude Code do not mind it.

Key model upgrades

Sottiaux said several recent model releases dramatically increased Codex usage, including:

GPT‑5.2 launched in December 2024, which improved Codex’s ability to handle long‑running tasks.

GPT‑5.5 released in April 2025, the first flagship AI since GPT‑4.5 (February 2024) to integrate a new base model.

Developers agree the recent releases boosted performance. DJ Sampath, senior vice president and general manager of Cisco AI Software and Platforms, said GPT‑5.5 handles long‑running tasks with far less human guidance.

“Since GPT‑5, Codex has consistently outperformed, but with GPT‑5.5 the gap is so large it’s impossible to ignore.” – Troyanovsky, Basis

Codex vs. Claude Code: developers’ real choice

Multiple developers noted that when engineers have a clear specification, Codex performs better; Claude Code is better for junior engineers or when the solution shape is uncertain.

Claude Code excels at predicting user intent even with vague descriptions.

“If I really don’t know what the solution looks like, Claude Code feels a bit better. When I have a precise spec, I trust Codex more.” – Varun Rau, software engineer at Notion

Embiricos added that updated models and the new Codex app allow Codex to understand ambiguous commands more effectively.

Rau also said Codex is “relentless” at fixing bugs. For example, when his team built a tool to launch new servers for Codex, they spent weeks tracking a bug that Codex automatically fixed each time the server started—something they only noticed after trying another coding agent.

Advantages can become drawbacks

Claude Code’s autonomous reasoning can sometimes lead it into dead‑ends or ignore developer instructions.

Codex’s “relentless” nature means it may chase many bugs even for tiny changes.

Desktop app vs. terminal: developers’ preference shift

When OpenAI launched the Codex desktop app in February, many developers switched to it, while they previously used the terminal (a plain‑text command window) for Claude Code.

Anthropic originally announced desktop Claude in November, but only redesigned the app in April.

Developers say the desktop approach is ideal for non‑technical users or those who want multiple coding agents to handle different tasks simultaneously. Max Schen, product lead at Notion, especially likes the ability to start coding tasks on a computer and then monitor progress via the newly released Codex mobile app.

Failure‑driven compensation: Codex team’s “tradition”

Codex leadership frequently resets usage limits when the app hits milestones or experiences outages.

“One of my principles is that if we mess up something, we must compensate users. That’s why whenever a failure occurs—whether it lasts five minutes, an hour, or four—we reset everyone’s rate limits.” – Sottiaux

Conclusion

OpenAI is embedding Codex into ChatGPT to push roughly 900 million consumer users into the emerging Agent era, while Anthropic’s $47 billion annualized revenue pressures OpenAI to move quickly. For developers, the battle means both Codex and Claude Code will keep evolving, usage caps will be reset, and tools will become increasingly useful.

The only awkward part for domestic users is that both Codex and Claude Code require overseas credit cards, stable foreign network access, and carry the risk of account bans. To bypass these hurdles, some turn to services like Code80, which convert real subscription accounts into API endpoints that work with Codex, Claude Code, and Claude Opus 4.8 without the need for overseas payment or network setups.

References

https://www.theinformation.com/articles/inside-openais-decision-combine-codex-chatgpt

FAQ

Q1: What is OpenAI’s “super app” and when will it be available?

A: OpenAI plans to combine ChatGPT, Codex and the Atlas browser into a desktop application that will launch in the next few weeks. In the short term users can choose whether Codex or ChatGPT answers a query; long term the model will decide automatically which to run and where.

Q2: How fast is Codex’s user growth?

A: By the end of May Codex surpassed 5 million weekly active users, up from 4 million in April and 3 million two weeks earlier—essentially doubling in two months. Enterprise revenue is growing 50 % week‑over‑week, and overall usage is up about 5 % daily.

Q3: How does Anthropic’s annualized revenue compare to OpenAI’s?

A: Anthropic’s annualized revenue recently topped $47 billion, five times its figure at the start of the year, driven mainly by Claude Code. OpenAI disclosed $25 billion in March and recent reports say it has now exceeded $30 billion.

Q4: Which should I choose, Codex or Claude Code?

A: Developers say use Codex when you have a clear specification; use Claude Code when the solution shape is uncertain. Codex is praised for precise execution and ruthless bug fixing, while Claude Code is better at anticipating intent for junior engineers.

Q5: Why does Codex now outperform ChatGPT?

A: The core reason is that Codex’s model and product are built by the same team, unlike OpenAI’s applications division, which is separate from the research organization. Codex also benefits from a stronger “tool‑calling framework,” the key technical reason for merging it with ChatGPT.

Q6: Is Codex cloud‑based or local, and how will it fuse with ChatGPT?

A: Codex runs locally on user devices and can access local files; ChatGPT runs in the cloud. This is the biggest integration challenge. Short‑term the plan is to let users manually pick Codex or ChatGPT; long‑term the model will decide automatically where to run.

Q7: How can users in China access Codex or Claude Code?

A: Both services require overseas credit cards and a stable foreign network, which creates barriers for domestic users. Some work around this by using services like Code80 that convert real subscription accounts into API endpoints, allowing direct use of Codex, Claude Code, or Claude Opus 4.8 without the usual payment or network hurdles.